sukju/cardiologent_dataset

CardioBench

Paired ECG + PPG waveforms with per-sample arrhythmia annotations, plus free-text reasoning QA, derived from VitalDB.

• 166,026 windows from 426 patients (train 302 / val 59 / test 65).
• Each window: time-aligned ECG and PPG (~10 s, 100 Hz) with a per-sample arrhythmia mask.
• Reasoning QA for training a waveform specialist and for evaluating diagnostic reasoning (T1/T2/T3).

Code (specialist training + diagnostic agent) lives in the companion GitHub repo.

Configs

Config	Rows	Description
waveforms	166,026	ecg / ppg / mask arrays per window
specialist_qa_ppg	166,026	per-window PPG diagnostic QA
specialist_qa_ecg	166,026	per-window ECG diagnostic QA
benchmark_reasoning	6,000	T1/T2/T3 reasoning QA (3,000 windows x both/ppg)

Waveforms

from datasets import load_dataset

ds = load_dataset("sukju/cardiologent_dataset", "waveforms", split="train")
ex = ds[0]
ex["window_id"]           # "1001_w0000"
ex["caseid"], ex["split"] # 1001, "train"
ecg, ppg, mask = ex["ecg"], ex["ppg"], ex["mask"]   # equal-length lists

Fields: window_id (str), caseid (int), split (train/val/test), ecg/ppg (list), mask (list).

ecg, ppg, mask share the same variable length L (one sample = 10 ms).

The mask — arrhythmia span and class

mask[i] labels sample i. The arrhythmia span is mask != 0; the class is the non-zero value on that span:

0 Normal   1 AF    2 SVT   3 VT    4 PVC
5 PAC      6 SND   7 MAT   8 AVB

PPG cannot separate some classes; for PPG-only evaluation merge VT,MAT -> VT/MAT and PVC,PAC -> PVC/PAC (7 classes). ECG uses all 9.

Reasoning QA

Specialist QA (specialist_qa_ppg / specialist_qa_ecg) — per-window diagnostic question + answer used to train the specialist:

{"id": "1001_w0000", "modality": "ppg", "question": "...", "answer": "..."}

Benchmark reasoning (benchmark_reasoning) — three reasoning probes per window, for the 9-class (both) and 7-class (ppg) settings:

window_id, setting, ecg_gt, ppg_gt,
T1_question/T1_answer,
T2_question/T2_answer,
T3_question/T3_answer

The three probes evaluate progressively deeper diagnostic reasoning:

• T1 — diagnosis & evidence. "Diagnose the rhythm and explain the key features (rate, R-R regularity, P-wave, QRS morphology) that support your diagnosis." Tests whether the model reaches the correct rhythm and cites the waveform evidence for it.
• T2 — differentials & exclusion. "List the plausible differential diagnoses and explain why you exclude each alternative (especially the most dangerous one)." Tests differential reasoning — naming the competing diagnoses and giving a reason each is ruled out.
• T3 — counterfactual pivot. "What single change in this recording would most likely change your diagnosis, and to what?" Tests understanding of the decision boundary — the one finding that would flip the call, and the resulting class.

Each record carries all three probes for one window under one setting; setting is both (9-class, ECG resolution) or ppg (7-class merged). ecg_gt / ppg_gt are the ground-truth classes for reference.

How the QA was generated

All free-text QA (both the specialist QA and the T1/T2/T3 probes) was generated with GPT-4o and then reviewed for quality. Generation was grounded so the answers stay faithful to the actual waveform rather than being free-form text:

1. Waveform image — the window is rendered with detected R-peaks marked, and passed to GPT-4o as an image.
2. Ground-truth anchor — the per-window arrhythmia label (from the mask) is supplied as a hint so the answer commits to the correct class. The anchor is a guide, not text to copy.
3. Quantitative features — deterministic signal measurements (heart rate, R-R variability, per-beat ΔRR and amplitude, premature-beat / pause markers) are provided so the reasoning cites real numbers.
4. Arrhythmia taxonomy — the class definitions above are given in the system prompt so terminology is consistent.

The model writes the diagnosis and its supporting reasoning from these inputs; the T1/T2/T3 probes additionally ask for differentials (T2) and the diagnosis-flipping counterfactual (T3). Outputs were then reviewed and low-quality or inconsistent items were regenerated. Ground-truth labels themselves come from the annotations, not from the model.

Source & license

Derived from VitalDB; please follow VitalDB's terms of use.